Year: 2012
Type: dissertation
Abstract: [ANGLÈS] A general framework for microarray classification relying on histogram based hierarchical clustering is proposed in this work. It produces precise and reliable classifiers based on a two-step approach. In the first step, the feature set is enhanced by histogram based features corresponding to each cluster produced via hierarchical clustering, where a parameter (maximum number of dominant genes) can be tuned based on the dataset characteristics. In the second step, a reliable classifier is built from a wrapper feature selection process called Improved Sequential Floating Forward Selection (IFFS) to properly choose a small feature set for the classification task. Considering the sample scarcity in the microarray datasets, a reliability parameter has been considered to improve the feature selection process along with classification error rate. Different combinations of error rate and reliability has been used as the scoring rule. Linear Discriminant Analysis (LDA) and K-Nearest Neighbour (KNN) classifiers have been used for this work and the performances has been compared. The potential of the proposed framework has been evaluated with three publicly available datasets : colon, lymphoma and leukaemia. The experimental results have confirmed the usefulness of the histogram based hierarchical clustering and the new representative feature generation algorithm. A gene level analysis has revealed that the best features selected by the feature selection algorithm has only very few basic constituent genes involved. The comparative results showed that the proposed framework can compete with state of the art alternatives.
Author: Sandeep Kottath
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W2137573828